Elnora AI Agent Hackathon Starter Kit

One command sets up everything you need to build AI agents on your laptop — your
coding agent (Claude Code or Codex, your choice) plus the dev tools that go
with it: Python, Node.js, Git, VS Code, and Obsidian.

Built for the Elnora AI agent hackathon: run one command, follow the
instructions, and get your first agents up and running.

Who this is for

Anyone who wants the fastest path from a fresh laptop to a working Claude Code or
Codex setup. It also works as a clean starting template for any agent project.

Requirements

• macOS or Windows 10/11 with admin rights (Homebrew or WinGet need them).
• A plan or API key for the agent you pick:
  • Claude Code — an active Claude Pro or Max
    subscription (or an Anthropic API key).
  • Codex — an active ChatGPT Plus/Pro plan (or an
    OpenAI API key).
• GitHub account, used in Phase 2 to create your private repo.

Install

macOS: open Terminal (Cmd+Space, type Terminal):

curl -fsSL https://raw.githubusercontent.com/Elnora-AI/elnora-ai-agent-hackathon-starter-kit/main/install.sh | bash

Windows: open PowerShell (Start key, type PowerShell). Not Command Prompt or cmd.

irm https://raw.githubusercontent.com/Elnora-AI/elnora-ai-agent-hackathon-starter-kit/main/install.ps1 | iex

A few things to expect:

• Before you start, sign into your accounts in a browser and leave the tabs
  open — your coding agent (Claude or ChatGPT for Codex) and GitHub. This
  confirms the accounts exist and makes the sign-in steps go through smoothly.
• It first asks which agent you want: Claude Code, Codex, or both. (Pick
  "both" and it installs both, then asks which one finishes setup right now.)
• macOS will prompt for your Mac login password (Homebrew). No characters appear as you type. Normal.
• A browser opens to sign into your agent (Claude Pro/Max, or ChatGPT for Codex). Sign in, then return to the terminal.
• Your agent takes over and finishes Phase 2 setup.

The script downloads installers from raw.githubusercontent.com, claude.ai,
and chatgpt.com over HTTPS without separate checksum verification. Running it
means trusting those sources.

If it stops, just run it again

Stopped, closed the terminal, or sign-in didn't go through? Run the same
install one-liner again (from Install above). It remembers the
workspace you already started and offers to resume it — already-installed
tools are skipped, and your .env / local config files are preserved. You'll
see Resuming where a previous run left off, so you know it's continuing,
not starting over.

The most common stopping point is the sign-in step — make sure your agent
account is active, then re-run and complete the login when the browser opens.

You can also run the setup script inside your workspace folder directly.
These commands work from any directory in a fresh terminal — replace
<your-workspace> with the workspace name you chose:

# macOS
bash ~/Documents/<your-workspace>/setup-mac.sh

# Windows
powershell -ExecutionPolicy Bypass -File "$env:USERPROFILE\Documents\<your-workspace>\setup-windows.ps1"

Want a completely clean run instead? Add --fresh to the end of either
command.

Script "not recognized" / not found? Setup already finished

If the setup script is missing from your workspace folder, that's not an
error — it means setup completed. The final cleanup step removes the
install scripts once everything is done, so there is nothing left to re-run.
To continue working with your agent, open a terminal and start it in your
workspace:

# macOS
cd ~/Documents/<your-workspace>
claude        # or codex

# Windows
cd $env:USERPROFILE\Documents\<your-workspace>
claude        # or codex

What happens

1. Phase 1 install: asks which agent you want (Claude Code / Codex / both)
   and a name for your workspace (used for both the local folder and the GitHub
   repo we'll create later), clones the repo to ~/Documents/<your-name>/, and
   installs your chosen agent(s) plus Homebrew/WinGet, Node.js, Git, Python, VS
   Code, GitHub CLI, and Obsidian. Existing installs are skipped. Output goes to
   ~/claude-starter-install.log.
2. Auth: your agent (required), GitHub CLI (skippable).
3. Phase 2 handoff: your agent verifies versions, creates your private
   GitHub repo, pushes the kit to it, runs a smoke test, and optionally sets up
   a knowledge base and the Vercel / Google Cloud integrations.

What gets installed

Phase 1 installs these on every run (skipping anything you already have):

In Phase 2, your agent offers to help you set up the following — it only
installs each one if you say yes:

Your agent can also set up a knowledge base (Obsidian vault — see the section
below) and finish enabling Chrome DevTools browser control in Phase 2 (the MCP
server is pre-connected; Phase 2 turns on Chrome's remote debugging — see
docs/chrome-devtools-mcp-setup.md).

What's already wired up

The kit comes pre-connected so these work on first launch:

• MCP servers: Chrome DevTools (control a browser), Context7 (live docs),
  grep (search public GitHub), Estonian (language tools).
• Vercel + v0 plugin: deploy and generate UIs with /vercel:deploy,
  /vercel:v0, and more.
• Vertex AI examples in examples/vertex/: generate images, video, and
  voiceover using your own Google Cloud project.

Repository layout

<your-workspace>/                          # e.g. carmen-agents, set during install
├── README.md                              # This file
├── INSTALL_FOR_AGENTS.md                  # Phase 2 sequence executed by your agent
├── RECOVERY.md                            # Failure modes and remediation steps
├── CLAUDE.md                              # Project instructions (Claude Code reads this)
├── AGENTS.md                              # Project instructions (Codex reads this)
├── TOOLS.md                               # Installed tools, plugins, and integrations
├── marketplace-plugins.md                 # Recommended plugin marketplaces
├── install.sh / install.ps1               # Bootstrap entry points
├── setup-mac.sh / setup-windows.ps1       # Phase 1 setup scripts
├── .env.template                          # API-key placeholders — copy to .env, never edit in place
├── .mcp.json                              # MCP servers (Chrome DevTools, Context7, grep, Estonian)
├── .gitignore
├── LICENSE                                # MIT
├── .claude/
│   ├── settings.json                      # Plugins, permissions, env defaults
│   └── knowledge-base.local.md.template   # Per-user knowledge-base config
├── plugins/                               # Bundled local plugins (vercel + v0)
│   └── vercel/
├── examples/
│   └── vertex/                            # Runnable image (nano-banana), video (Veo 3) + voiceover scripts
└── docs/
    ├── getting-started.md                 # Daily-workflow guide + manual fallback
    └── google-cloud-vertex-setup.md       # gcloud + Vertex AI: image, video, voiceover + more

Post-install

You'll have a private GitHub repo on your account. The origin remote
points at it (verify with git remote -v). From here it's yours: commit,
push, branch, rename. Keep this one private; spin up a separate public repo
if you want to share something later.

Reopening the project in VS Code: setup leaves a
<your-workspace>.code-workspace file in the folder and opens that (not the
bare folder), so your project shows up as one clear entry under
File → Open Recent and can be pinned to the Dock/taskbar. Setup also turns
on VS Code's "reopen windows on launch" so quitting and relaunching brings the
project back automatically — no hunting for the folder.

Knowledge base (optional)

The kit can connect Claude to a local Obsidian vault or any directory.
Phase 2 auto-detects vaults in iCloud, Google Drive, OneDrive, Dropbox, and
~/Documents. Config lives in .claude/knowledge-base.local.md
(gitignored). See CLAUDE.md → "Knowledge Base".

Troubleshooting

See RECOVERY.md. For unresolved issues, attach
~/claude-starter-install.log.

Manual setup

No Claude Pro/Max, or the install failed? docs/getting-started.md walks
through the equivalent manual steps.

Configuration

Defaults in .claude/settings.json:

• CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 enables multi-agent teams.
• CLAUDE_CODE_NO_FLICKER=1 opts into the full-screen alt-buffer renderer.
• autoUpdatesChannel: "latest" opts into auto-updates on the latest channel. Use "stable" for ~1-week-old builds. Ignored for package-manager installs.
• remoteControlAtStartup: true auto-enables Remote Control. Sessions are reachable from any device signed into your Claude account; review before enabling on machines that handle proprietary data.
• enableAllProjectMcpServers: true auto-approves every MCP server in .mcp.json (Chrome DevTools, Context7, grep, Estonian) so they're connected and usable on first launch — no per-server approval prompt. Remove this key if you'd rather approve each server manually.

Set values to false or "0" to disable.

License

MIT.